Reproduce “NFLAT: Non-Flat-Lattice Transformer for Chinese Named Entity Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shuang Wu, Xiaoning Song, Zhenhua Feng, Xiao-Jun Wu
Researchers from Jiangnan University and the University of Surrey developed NFLAT, a Transformer-based model for Chinese Named Entity Recognition that efficiently integrates lexical information. This approach utilizes a non-flat-lattice inter-attention mechanism to fuse character and word sequences, achieving state-of-the-art F1 scores across multiple benchmarks while significantly reducing computational and memory usage compared to previous methods like FLAT.
Key insights
- Decoupling the lexicon fusion process from the main contextual encoding via inter-attention effectively mitigates the quadratic computational and memory costs associated with flat-lattice architectures.
- Relative position encoding, accounting for character-word head and tail offsets, is crucial for precisely integrating word boundary information into character representations through inter-attention.
- An `<non_word>` tag in the lexicon matching process enables auto-discovery of meaningful words and auto-suppression of noise (e.g., punctuation, unmatched single characters), enhancing model robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.